Download How External Factors Affect Domestic Economy

WP/15/269
How External Factors Affect Domestic Economy:
Nowcasting an Emerging Market
By Serhat Solmaz and Marzie Taheri Sanjani
© 2015 International Monetary Fund
WP/15/269
IMF Working Paper
Strategy, Policy, and Review Department
How External Factors Affect Domestic Economy: Nowcasting an Emerging
Market
Prepared by Serhat Solmaz and Marzie Taheri Sanjani1
Authorized for distribution by Vikram Haksar
December 2015
IMF Working Papers describe research in progress by the author(s) and are published to
elicit comments and to encourage debate. The views expressed in IMF Working Papers are
those of the author(s) and do not necessarily represent the views of the IMF, its Executive Board,
or IMF management.
Abstract
External headwinds, together with domestic vulnerabilities, have loomed over the prospects of
emerging markets in recent years. We propose an empirical toolbox to quantify the impact of external
macro-financial shocks on domestic economies in parsimonious way. Our model is a Bayesian VAR
consisting of two blocks representing home and foreign factors, which is particularly useful for small
open economies. By exploiting the mixed-frequency nature of the model, we show how the toolbox
can be used for “nowcasting” the output growth. The conditional forecast results illustrate that regular
updates of external information, as well as domestic leading indicators, would significantly enhance
the accuracy of forecasts. Moreover, the analysis of variance decompositions shows that external
shocks are important drivers of the domestic business cycle.
JEL Classification Numbers: C11, C53
Keywords: Nowcasting, Bayesian estimation, VAR, Emerging Market
Author’s E-Mail Address: [email protected], [email protected]
1
The authors would like to thank Michal Andrle, Gregorio Impavido, Helene Poirson Ward, Mika Kortelainen,
Ernesto Ramirez, Antonio Spilimbergo, Alexander Tieman, Ivanna Vladkova Hollar, and Kevin Wiseman for
useful comments.
3
Table of Contents
Abstract ..................................................................................................................................... 2
I. Introduction ........................................................................................................................... 4
II. Model and Data .................................................................................................................... 6
III. Impulse Response Analysis ................................................................................................ 9
IV. Variance Decomposition .................................................................................................. 10
A. Forecast Error Variance Decompositions (FEVD) ........................................................ 10
B. Historical Variance Decompositions .............................................................................. 12
V. Counterfactual Simulations ................................................................................................ 14
VI. Conditional Forecast ......................................................................................................... 16
VII. Conclusion ....................................................................................................................... 18
VIII. References ...................................................................................................................... 19
IX. Appendix .......................................................................................................................... 20
A. Data ............................................................................................................................ 20
X. Structural Forecast Error Variance Decompositions ......................................................... 21
XI. Nowcasting in Practice ..................................................................................................... 22
Figures
1. External Shocks and Responses of Domestic Variables .................................................... 10 2. Forecast Error Variance Decompositions of Home and Foreign Shocks ........................... 11 3. Historical Variance Decompositions .................................................................................. 13 4. Counterfactual Simulations ................................................................................................. 15 5. RMSEs for Different Time Vintages (quarterly nowcast of output growth) ...................... 17 6. Stylized Facts: Model Variables (HP filtered) .................................................................... 21 7. Structural Forecast Error Variance Decompositions .......................................................... 22 8. GDP Growth (% YoY) and Nowcast .................................................................................. 23 Tables
1. Out of Sample Forecast Errors for Different Time Vintages .............................................. 18
2. Out of Sample RMSEs for Different Time Vintages .......................................................... 18
3. Data Description and Sources ............................................................................................. 20
4. Cross-Correlation of Output Gap with following Variables ............................................... 20
5. Calendar of Turkey’s GDP Releases .................................................................................. 23
4
I.
INTRODUCTION
Emerging market (EM) vulnerabilities to external headwinds and their intrinsic
domestic fundamentals have returned to the center of policy discussions. Post global
financial crisis (GFC) spillover of advanced economies’ policies, along with various
uncertainty factors, tangled with weak domestic fundamentals, loom over emerging markets
with large trade links and high exposure to the global financial cycle.
This paper proposes a methodology to quantify the impact of external and domestic
macro-financial impulses in explaining the dynamic of home economy. Our proposed
framework allows for both forward looking and historical analysis. Historical analysis sheds
light on the roles of both external and domestic conditions in explaining the business cycle in
EMs, while the forward-looking analysis can guide policy by measuring the impact of new
information on forecasts. By limiting the model to key explanatory variables, we aim to focus
on the main channels of propagation and spillover.
Our methodology is based on a Mixed-Frequency Bayesian VAR (MFBVAR), which
consists of two blocks, home and foreign. While the home economy can be any open
economy exposed to global factors through trade and financial channels, we have selected
Turkey as the home economy for our methodology. The mixed-frequency feature of our
model allows us to combine monthly and quarterly information, which we use to compute
our conditional forecast, or “nowcast.”
To illustrate the application of our proposed methodology, we apply it to the case of
Turkey. Turkey, a very open economy that exports to major advanced economies, especially
Europe,2 is completely dependent on the importation of primary commodities, and has major
trade links with many countries in its region. Moreover, financial factors such as risk taking
2
While Turkey’s trade links with Europe have been traditionally large and important, Turkey has various
trading partners and their relative share change over time. For example in 2013 Iraq was the Turkey’s 2nd
biggest export market.
5
channel and capital flow are also important. These characteristics make it an ideal test case.
This toolbox can be used for other small open economies. To do so the practitioner needs to
identify the relevant external variables to be included in the model, by assessing the statistics
concerning the main trading and financial trading partners, the degree of trade and financial
openness and the reliance on oil and other commodities. Table 4 in appendix presents
descriptive statistics of the variables in the model and their interdependence. We try to keep
the model minimal, in order to make it easy to use.
The choices of leading indicators and relevant global factors are guided by theory and
application in order to describe the evolution of the domestic economy properly.
Application dictates the choice of the variables in the model. In our case, the framework is
designed for GDP nowcasting.3 The leading indicators in the home economy should have
thorough information content about investment, trade, and financial activities. The external
factors that form the foreign block of our model are chosen to reflect the macroeconomic
state of the main trading and financial partners of the home economy. The home block
consists of five variables, while our foreign block consists of only three variables. The
information content of global oil prices, US corporate risk spread, and US industrial
production, can signal shocks related to global demand, finance, and trade.
We find that external shocks can account for more than 50% of the GDP growth
dynamic in Turkey. Historical variance decompositions illustrate that a decrease in the risktaking appetite in the US or strong US industrial production (IP) provide positive impulses
for our emerging market economy’s output growth. Our counterfactual simulations show that
external shocks play a key role in explaining the recession and subsequent rebound. By
exploiting the mixed-frequency nature of the model, nowcasting can be used to quantify the
direction of macroeconomic developments, as the high frequency data become available. We
should note that our model improves upon the benchmark naïve model.
3
For example, for an inflation forecasting exercise we need different indicators than a GDP forecasting case.
6
The paper is organized as following. Section 2 discusses the model’s specifications and
database. Sections 3 to 6 present the results, namely: impulse response function, variance
decomposition analysis, counterfactual simulations, and conditional forecast results,
respectively. Section 7 concludes. References and Appendices are in Sections 8 through 11.
II.
1.
MODEL AND DATA
We use a mixed-frequency Bayesian Vector Autoregression (VAR) model for our
analysis, consisting of two blocks: a home economy and a foreign economy, See,
Österholm and Zettelmeyer (2008); Abrego and Österholm (2008); and Andrle, Ho and
Saltos (2013). The VAR model is an effective tool for studying complex, dynamic
interrelation in data. The empirical analysis of this model is four-fold. First, by estimating the
impulse response functions using VAR, we can trace out the effects of policy, innovations
from external factors, and risk on the economy. Second, we analyze historical variance
decompositions to demonstrate the impact of external and domestic shocks in explaining the
volatility of the domestic economy. Third, we compute the in-sample conditional forecasts to
assess the information-optimality of the external and domestic factors independently. Finally,
we provide a nowcast of GDP growth using monthly information releases. In what follows,
we describe the model specification.
The model is a structural VAR with only one lag, due to the short sample size and in order to
avoid the over parameterization problem. All the variables in the model are stationary and
have a zero mean, since we demeaned and detrended all the data using the Hodrick-Prescott
(HP) filter. The model converges to a steady state. With Y(t) being the vector of endogenous
variables with n entries, assume the economy evolves as:
ᵼ
where the constant term
is a
ᵣ
Ɛᵼ, Ɛ ~ . .
0,
,
x1 vector, but in our case it is zero since the data is
demeaned, where v is the number of variables, ᵣ is a x
autoregressive matrix, and Ɛᵼ is a
7
x1 white noise process with covariance matrix
. Since we demean the model, the closed-
form notation of the model is:
ᵼ
1
Ɛᵼ, Ɛ ~ . .
0,
.
The identification assumption ensures no spillover from the domestic economy to the
external block while, the external block does affect the domestic economy. We estimate
the model with the standard OLS estimates by imposing block erogeneity condition. We use
a multi-step forecasting error approach to evaluate the forecasting power of our model by
recursively computing the h-step ahead forecast, see Marcellino, Stock, and Watson (2006).
We analyze the period after the banking crisis of 2001 in Turkey. The data cover the period
2003Q2 to 2014Q1. The number of explanatory variables in the model is kept to a minimum,
so that we can track through the main channels of propagations. The information content of
variables in the model signals macro-financial developments in aggregate output.
Specifically, we use only the price of oil, US industrial production,4 and US risk spread
(defined as the spread between Moody’s BAA and AAA corporate bond rates) in the foreign
economy block. Intuitively, global oil price and US IP reflect the dynamic of commodity
cycle and demand pressure respectively; US risk spread signal foreign investors risk taking
appetite. Note that there are various risk indicators such as EMBI, VIX, FSI (financial stress
index), which can be used for other small open economies, nevertheless for the case of
Turkey we found that US risk spread suits the nowcasting application better.5
Generally to tailor the model for other economies’ nowcasting, the practitioner should choose
the domestic variables by analyzing all possible soft and hard leading indicators for which
data is available. For example, other potential candidates include consumer and business
confidence indicators, unemployment, consumption expenditure, manufacturing trade sales,
4
Hard data such as industrial production or imports carries a lot of information about the GDP growth
in Turkey (See Akkoyun and Gunay, 2012).
5
One way of comparing the explanatory power of various indicators is to check mean square forecast
error.
8
stock price index, capital utilization and exchange rate. In choosing the external variables, the
real and financial variables of the main trading partners are the key.
We have examined a panel of High Frequency Indicators (HFIs) in order to choose domestic
leading indicators. In the domestic economy block we use real GDP, industrial production,
three-month weighted average interest rates, imports, and loans to the private sector. Loans to
the private sector are a particularly informative leading indicator for Turkey, in the context of
a small open economy, since it contains relevant information about capital inflows and credit
availability. IP reflects production performance, import can shed light on the trade dynamic,
and interest rate present saving-investment dynamic. Variables are seasonally adjusted by
official sources and authors’ computations (X12 ARIMA standard procedure).
Data Y
In the domestic block, we order slow-moving variables6 first so that they can
contemporaneously affect fast moving variables.7 Cholesky decomposition is used to
6
Note that GDP is quarterly and the rest of the variables are monthly.
9
estimate an orthogonal covariance matrix of the residuals. The ordering of the variables in
the model, our vector of data, is specified as follows. Our identification implies that external
shocks have an impact on domestic variables.8 Moreover, in the domestic block GDP is
presumed to be uncorrelated with other (non-GDP) domestic shock processes.
To make the series stationary, we compute cyclical components of the data. These are
obtained by applying the HP filter to 100 times the natural log of data. We apply the HP filter
only to the level of domestic economy interest rates and US financial risk spread data. Since
we use leading indicators in this forecasting exercise, the cyclical components are quite
crucial for signaling the drivers of forecast in rapid growth, stable, and slowdown periods.
Hence, calculations based on the cyclical component improve the forecast performance of the
model.
III.
IMPULSE RESPONSE ANALYSIS
By comparing the impulse responses of GDP growth to external shocks, one can observe
that oil shock can lead to the largest drop in output growth.
Figure 1 reports the median of the posterior distribution of the impulse responses to adverse
shocks of risk spread, US IP, and global oil price. The spread shocks are financial shocks,
affecting the supply of funds, while oil price and US IP shocks are demand shocks. Upon the
impact of a one-standard-deviation oil shock (spread shock and US IP), output drops 1.3 (0.7
and 1) percentage points relative to its steady state. IP, credit, and imports commove with
output growth. The interest rate increases in response to adverse supply shocks, reflecting the
tradeoff between output and inflation stabilization objectives, while it decreases in response
to the demand shocks.
7
Giannone et. al (2015) also place GDP as the first variables. We have done some robustness check by
changing the ordering of the variables in the domestic block and we found that the main finding of the paper
stays qualitatively the same.
8
This identification will not be applicable for US economy since it is a large open economy with complex links
to the global economy.
10
Figure 1. External Shocks and Responses of Domestic Variables
IV.
VARIANCE DECOMPOSITION
A. Forecast Error Variance Decompositions (FEVD)
By using variance decomposition analysis, we quantify the impact of external shocks on
the domestic economy at different horizon. This exercise demonstrates that global factors
have significant impacts on our small open emerging economy, which could further have
important policy implications. By computing variance decomposition we essentially project
the dynamic of an endogenous variable into the different shocks component of the VAR, in
order to measure the information content of relative importance of each shock in explain the
variables in the VAR. We can rewrite the model using the lag operator as following:
ᵼ
Ɛᵼ → Y
I
AL
Ɛᵼ
and the error in forecasting Y(t) for each horizon h is:
B L Ɛᵼ,(3)
11
Y h
B L Ɛ
,(4)
Hence, the variance of the total forecasting error is given by:
Y h
B L ΣB L ,(5)
We can compute individual shock’s variance decomposition as the individual share of the
total variance decomposition. To compare the domestic versus external shocks, we aggregate
FEVD of all the domestic shocks together and call it “home”; we do the same for the external
shocks and we call them “foreign”. Figure 2 shows total forecast error variance
decomposition over 10 quarters following aggregated shocks to the home economy. The blue
bars show the aggregate contribution of the domestic shocks, while the white bars explain the
aggregated external shocks. All shocks in the model are considered unanticipated. The large
share of white bars in the panel suggests that most of the volatility in the domestic indicators
is explained by external factors.
Figure 2. Forecast Error Variance Decompositions of Home and Foreign Shocks
12
Note that, by construction, our identification assumption implies that there is no spillover
from the domestic economy to the external variables; therefore, the FEVDs of external
variables is merely explained by the shocks originating in the foreign block. By investigating
the panel, one can draw the following observations. First, the impact of domestic shocks in
the first quarter is the strongest. Second, more than 60% of output FEVD is driven by
external factors. While the response of the loans to private sector is determined largely by
domestic factors in the short term, in the medium-to-long-term the share of external shocks
increases. This is due mainly to capital flows that are converted to credit in the home
economy. Similarly, the response of short-term interest rates is explained mainly by the
domestic factors, as the monetary policy responds to business cycle shocks.
Figure 7 in the Appendix shows the individual variables’ structural historical variance
decompositions. The large impact of external variables on domestic economy is not
surprising, as Turkey has become increasingly open economy after 2001. Both pull and push
factors, global financial cycle, appreciated currency, and improved domestic prospect, played
role in increasing capital inflow into the economy.
B. Historical Variance Decompositions
Drivers of the domestic business cycle can be determined by computing the dynamic
contribution of each shock in explaining the historical path of the variables. Figure 3
shows the year-on-year, historical variance decompositions (HVD) of the variables in the
model. Intuitively given a variable, such as GDP, we project the variable in the orthogonal
space of the shocks; hence, each colored bar represents the contribution of the related shock
in explaining the dynamic of the variable of interest. Such quantitative analysis of business
cycle drivers can help tailor the policy mix. For example, in episodes of high external
spillover in the domestic economy, identifying whether the spillover is uncertainty driven or
due to change in the fundamentals can help domestic policy makers choose their policy
instruments more effectively.
13
Figure 3. Historical Variance Decompositions
The GDP dynamic shows a large degree of persistence between 2011 and 2013. The external
channels, both real and financial, have been very important during the turmoil of 2012-2013.
A combination of Euro Area double-dip recession, uncertainty around the US fiscal cliff,
Turkey’s political unrest, and other factors led to an increase in volatility and market
perception of risk in emerging markets.9 Hence, US IP is playing a negative but small role,
while US risk spread is playing a large and positive role in explaining the domestic
economy’s business cycle between 2012 and mid-2013. This is intuitive, as US IP is a proxy
9
With large declines in portfolio investment, gross capital inflows to central and southeastern Europe turned
sharply negative in the third quarter of 2013 and dropped substantially for Turkey (Figure 2.4, WEO April
2014, page 57).
14
for external real activity, and US risk spread is a proxy for the risk-taking appetite of foreign
investors. Hence, the contribution of the US IP has an opposite effect compared to the
contribution of the risk spread, highlighting the countercyclical nature of the risk spread.10
Finally, one can observe that the real GDP chart in the top left hand-side commoves with the
IP in the top middle chart.11
Domestic factors have a minor role in explaining the dynamic of short-term rates. Moreover,
in 2012 the contribution of risk spread has been positive and significant, but the explanatory
role of the spread is diminishing in the latter part of the sample. Credit commoves with GDP,
but nonetheless, the role of domestic factors is decreasing toward the end of the sample.
V.
COUNTERFACTUAL SIMULATIONS
How important are external shocks in explaining historical data? We answer this
question by computing three counterfactual economies: (i) a simulated economy with all the
shocks included; (ii) a simulated economy where we exclude domestic shocks, but the
external shocks still affect the model dynamic; and finally, (iii) a simulated economy where
we exclude external shocks, while the domestic shocks are still operating. To compute these
counterfactual economies we first estimate the parameters and historical paths of shocks of
the model using the whole information set; then we simulate our BVAR model using a
subsample of the shocks by setting zero coefficients on the selected residuals, specifically,
external shocks (ii) and domestic shocks (iii). The simulated paths now correspond to a
hypothetical economy with no forecast errors in the residuals of the foreign block and
domestic block, respectively.
Figure 4 presents a panel of the domestic variables’ historical paths under the above three
scenarios. The red curves denote the baseline economy in the presence of all the shocks. The
10
When US economy is growing, foreign investors’ home bias increases; as a result FDI to emerging
economies shrinks, which further affects credit dynamic in domestic economy. The increase in US corporate
risk spread has an opposite effect.
11
This is mainly due to the methodology used by Turkish Statistical Institute (TurkStat) to construct the Real
GDP, which assigns a high weight on IP.
15
blue curves illustrate our counterfactual economy in which we turn off the external shocks.
Finally, the green curves show the counterfactual economy without domestic shocks. By
comparing the three curves, one can assess the information content of different shocks in
explaining the historical paths of the variables in the economy.
In the wake of the recent global financial crisis, the Turkish economy experienced a
recession and a subsequent rebound in 2009 and 2010, partially due to the negative spillover
from the US economy. The periods before and after the 2009 recession are considered
normal times in this paper. By comparing the green and blue curves in the
Figure 4, one can observe that, while domestic shocks play an important role in explaining
the business cycle dynamic during normal times, the external shocks can explain the joint
dynamic of downturn and rebound better. A model that does not account for external shocks
predicts a shorter duration for the 2009 downturn episode and a too-speedy recovery.
Figure 4. Counterfactual Simulations
16
VI.
CONDITIONAL FORECAST
The mixed-frequency nature of our model allows for using the new release of
information in updating the forecast of the GDP. This is particularly useful because GDP
has a slow release and is available on a quarterly basis, but it can be estimated using higherfrequency variables. This work belongs to the strand of literature that studies “nowcasting
GDP.” Nowcasting is an exercise for computing the current-quarter forecast of key indicators
of real economic activity, namely GDP, using the information flow. Any release of high
frequency data might convey economically significant information, which has an impact on
GDP. Following Giannone, Reichlin, and Small(2008), the conditional forecast problem is
defined as following:
Ε
∆
Ε
Δ
Ε δ
,
,
,…
,
where the Ε
Δ
Ε months.
Every
δ
(6)
(7)
,
,
,(8)
shows the nowcast of quarterly GDP at period t, conditional
on the information set available up to month M+1, Δ
for
month,
with
new
data
. Q stands for quarters and M stands
available,
the
forecast
is
revised
, using the Kalman filter to account for the missing observations.
Nowcasting improves as new information becomes available throughout the month. We
evaluate the forecasting power of the model by computing one-quarter ahead root-meansquare errors (RMSEs) over different horizons and different vintages. The RMSEs of the
model’s forecast for GDP growth of Turkey are calculated as:
∑
, (9)
17
where
1
6.
Figure 5 summarizes the RMSE results. The blue bars are unconditional forecasts; the red,
green, and yellow bars correspond to the nowcast conditional on the data available up to the
first, second, and third months of each quarter; and finally, the yellow bars correspond to the
nowcast conditional on the data available up to the third month of each quarter.
Figure 5. RMSEs for Different Time Vintages (quarterly nowcast of output growth)
We also compared our annual forecasts with the World Economic Outlook (WEO) forecast
for the recent four years, including the 2009 recession and 2010 recovery. To compute the
out-of-sample forecast for any given year i, we restrict ourselves to the data available up to
the previous year i-1.12
Table 1 and
Table 2 show the comparison between out-of-sample annual (four steps ahead) forecast
errors and RMSEs of our model with the spring WEO forecast. Forecast errors are calculated
as the difference between the forecast and actual. For example, Table 1 shows that WEO’s
12
We use final estimate for the year instead of initial estimates.
18
forecast of 2009 GDP growth has been 4% above the actual, while our model predicted the
2009 recession with a somewhat bigger trough. Althoguh using a VAR framework with
domestic leading indicators and global variables provides us with a parsimonious forecasting
framework, with a good accuracy, practitioner should be careful relaying only on single
forecasting model. Usually combination of several forecasting techniques should be used to
more meaningfully understand the future developments in the economy.
Table 1. Out of Sample Forecast Errors for Different Time Vintages
Forecast Turkey’s output Growth
2009
2010
2011
2012
WEO
4.0
4.2
-0.2
0.6
Our Model
-2.3
2.3
0.4
-0.1
Table 2. Out of Sample RMSEs for Different Time Vintages
Forecast Turkey’s output Growth
2009-2012
2010-2012
WEO
2.9
2.4
Our Model
1.6
1.3
VII.
CONCLUSION
This paper proposes a nowcasting toolbox for output growth by specifying an explicit role for
global factors. By presenting impulse response functions, variance decompositions, and
counterfactual simulations, we illustrate how external shocks, together with domestic shocks,
can explain the dynamic of variables in the economy. This empirical setup parsimoniously
combines the information content of domestic and external leading indicators; hence, it is
suitable for any small open economy that has large trade and financial links to advanced
economies. To illustrate the application of our proposed methodology, we apply it to the case
of Turkey. Our conditional forecast results suggest an improvement in forecasting power, as
we not only exploit the mixed-frequency nature of the model, we can also measure the
impulses from external factors.
19
VIII.
REFERENCES
Abrego, L., and O. Par, 2008, “External Linkages and Economic Growth in Colombia: Insights from
a Bayesian VAR Model,” Working Paper WP/08/46 (Washington: International Monetary Fund).
Akkoyun,H. C. , and M. Gunay, 2012, “Nowcasting Turkish GDP Growth,” Working Paper,
12/33,Central Bank of Republic of Turkey.
Andrle, M., R. Garcia-Saltos, and G. Ho, 2013, “The Role of Domestic and External Shocks in
Poland: Results from an Agnostic Estimation Procedure,” Working Paper WP/13/220 (Washington:
International Monetary Fund).
Banbura, M., D. Giannone, and L. Reichlin, 2010, “Large Bayesian VARs,” Journal of Applied
Econometrics, 25(1), 71–92.
Benes, J., M. K. Johnston, and S. Plotnikov, IRIS Toolbox Release 2014/06/19 (Macroeconomic
modeling toolbox), software available at http://www.iris-toolbox.com.
Giannone, D., M. Lenza, and G. E. Primiceri, 2015, “Prior Selection for Vector Autoregressions,”
The Review of Economics and Statistics, May 2015, 97(2): 436–451.
Giannone, Domenico, Lucrezia Reichlin, and David Small, 2008, "Nowcasting: The real-time
informational content of macroeconomic data," Journal of Monetary Economics, Elsevier, vol. 55(4),
pages 665-676, May.
Kadiyala, K. R., and S. Karlsson, 1997, “Numerical Methods for Estimation and Inference in
Bayesian VAR-Models,” Journal of Applied Econometrics, 12(2), 99–132.
Koop, G., and D. Korobilis, 2010, “Bayesian Multivariate Time Series Method for Empirical
Macroeconomics,” Foundations and Trends in Econometrics, 3(4), 267–358.
Litterman, R. B., 1986, “Forecasting with Bayesian Vector Autoregressions – Five Years of
Experience,” Journal of Business and Economic Statistics, Vol. 4, pp. 25–38.
Marcellino,M., J. H. Stock, and M. W. Watson, 2005, “A Comparison of Direct and Iterated
Multistep AR Methods for Forecasting Macroeconomic Time Series,” Journal of Econometrics, 135 ,
499–526.
Österholm, P., and J. Zettelmeyer, 2007, “The Effect of External Conditions on Growth in Latin
America,” IMF Working Paper WP/07/176, International Monetary Fund.
20
IX.
Appendix
A. Data
Table 3. Data Description and Sources
Variable
Source
Gross Domestic Product (SWDA, Thous.98.TL)
TurkStat
Industrial Production: Total Industry (SA, 2010=100)
TurkStat
Foreign Trade: Total Merchandise Imports, c.i.f. (SA, Mil.US$)
TurkStat
Deposit Money Banks Loans: Private Sector (EOP, SA, Thous.TL)
Central Bank of the Republic of
Turkey
Weighted
Average
Interest
Rates
for
TL
Deposits
{Ex
Central Bank of the Republic of
Sight/Interbank}: Up to 3Mo (%)
Turkey
World Crude Oil {Petro}: Brent,WTI,Dubai Fateh Spot Price Avg.(
IMF
US$/Bbl)
US: Industrial Production excluding Construction (SA, 2007=100)
Federal Reserve Board
Risk Spread
Baa-Aaa
Moody's Seasoned Baa Corporate Bond Yield (% p.a.)
Moody's
Moody's Seasoned Aaa Corporate Bond Yield (% p.a.)
Moody's
Table 4. Cross-Correlation of Output Gap with following Variables
(Cyclical component of series (HP filtered))
Output GAP
Domestic IP
Short-term Rate
Import
Credits
Nominal POIL
US IP
US Financial Riskness
t-3
0.33
0.3
0.5
0.39
0.69
0.26
0.62
0.38
t-2
0.56
0.55
0.4
0.64
0.83
0.51
0.8
0.05
t-1
0.82
0.79
0.21
0.83
0.87
0.7
0.9
-0.35
t
1
0.94
-0.12
0.89
0.75
0.79
0.85
-0.68
t+1
0.82
0.76
-0.29
0.62
0.52
0.56
0.67
-0.77
t+2
0.56
0.51
-0.38
0.23
0.31
0.13
0.44
-0.61
t+3
0.33
0.22
-0.38
-0.03
0.15
-0.13
0.19
-0.52
21
Figure 6. Stylized Facts: Model Variables (HP filtered)
X.
STRUCTURAL FORECAST ERROR VARIANCE DECOMPOSITIONS
Figure 7, below, shows the individual variables’ structural forecast error variance
decompositions. The rows illustrate the indicators, and the columns illustrate the shocks.
Each chart shows the relative contributions of residuals. By investigating the variance
decomposition of the Real GDP on the top row, one can highlight the important role of
global risk factors, captured by the price of oil, and US investment risks, captured by the
spread, through the capital flows and foreign financing channels.
22
Figure 7. Structural Forecast Error Variance Decompositions
XI.
NOWCASTING IN PRACTICE
In this section, we give an example of how we use monthly information in producing a
nowcast of GDP within each quarter. We assume that the Turkish GDP is available up to
2014Q2; this exercise aims at computing a point forecast of 2014Q3. The domestic variables,
except short-term interest rates, are available up to 2014M7, while external variables are
available up to 2014M8. As new monthly information is published, we update our estimate of
the variance-covariance matrix and update the nowcasts. Our results support the findings in
the literature regarding the importance of both IP and imports as leading indicators of
Turkish aggregate output. In addition, we find that loans to the private sector have useful
23
information as a leading indicator. We are using a conditional forecast to parsimoniously
combine high frequency information in nowcasting real activities.
Table 5. Calendar of Turkey’s GDP Releases
Month
Release Date
Reference Quarter
March/April
Last week of March or 1st week of April
Q4
June
After first week
Q1
September
2nd week
Q2
December
2nd week
Q3
Therefore, nowcasting will be implemented using available information based on Figure 8.
Conditional forecasting on M1 includes information of first month of quarter for all the
monthly variables. Conditional forecasting on M2 includes information of second month of
quarter only for short-term interest rate and external variables in the model.
Figure 8. GDP Growth (% YoY) and Nowcast